config.json

{
  "allow_neg_eigval": false,
  "architectures": [
    "QuasarForCausalLM"
  ],
  "attn_mode": "chunk",
  "auto_map": {
    "AutoConfig": "configuration_quasar.QuasarConfig",
    "AutoModelForCausalLM": "modeling_quasar.QuasarForCausalLM"
  },
  "bigmac_r": 0.25,
  "bos_token_id": null,
  "conv_bias": false,
  "conv_size": 4,
  "d_ff": 4096,
  "d_model": 1536,
  "dense_input_layers": 4,
  "dropout": 0.0,
  "dtype": "bfloat16",
  "eos_token_id": 248044,
  "expand_k": 0.5,
  "expand_v": 1.0,
  "fuse_cross_entropy": true,
  "fuse_norm": true,
  "fuse_swiglu": true,
  "gated_layers": 2,
  "gla_mode": "chunk",
  "gradient_checkpointing": false,
  "head_dim": 128,
  "hidden_act": "silu",
  "hidden_ratio": 4,
  "hidden_size": 1536,
  "hybrid_layer_types": [
    "quasar",
    "quasar",
    "quasar",
    "quasar",
    "gla",
    "gla",
    "quasar",
    "quasar",
    "quasar",
    "quasar",
    "gla",
    "gla",
    "quasar",
    "quasar",
    "quasar",
    "quasar",
    "gla",
    "gla",
    "quasar",
    "quasar",
    "quasar",
    "quasar",
    "gla",
    "gla"
  ],
  "initializer_range": 0.02,
  "intermediate_size": 4096,
  "layer_types": [
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention",
    "linear_attention"
  ],
  "looped_injection_init": 0.1,
  "max_position_embeddings": 16384,
  "max_seq_len": 16384,
  "memory_dim": 128,
  "memory_slots": 128,
  "model_type": "quasar",
  "moe_aux_loss_coeff": 0.0001,
  "moe_type": "bigmac",
  "moe_z_loss_coeff": 0.0001,
  "n_heads": 12,
  "n_layers": 24,
  "norm_eps": 1e-06,
  "num_attention_heads": 12,
  "num_heads": 12,
  "num_hidden_layers": 24,
  "num_key_value_heads": 12,
  "num_loops": 1,
  "num_routed_experts": 64,
  "num_shared_experts": 1,
  "num_v_heads": null,
  "pad_token_id": 248044,
  "quasar_layers": 4,
  "residual_scale": 0.1,
  "rms_norm_eps": 1e-06,
  "rope_theta": 1000000.0,
  "routed_expert_size": 256,
  "shared_expert_size": 3072,
  "smebu_beta": 0.5,
  "smebu_kappa": 2.0,
  "smebu_lambda": 0.002,
  "tie_word_embeddings": false,
  "top_k": 4,
  "transformers_version": "5.9.0",
  "use_cache": false,
  "use_gla_first": false,
  "use_l2warp": false,
  "use_looped_injection": false,
  "use_short_conv": true,
  "vocab_size": 248320
}